Research map / 3D vision / 3D generation
SG2Loc: Sequential Visual Localization on 3D Scene Graphs
Nicole Damblon, Olga Vysotska, Federico Tombari, Marc Pollefeys, Daniel Barath
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add depth score to semantic score (ablation) | SG2Loc · 3RScan first 4 scenes, 5 frames | R@25cm,10deg: 0.35 -> 0.54 | Improved |
| add RGB (SSIM) score on top of semantic+depth (ablation) | SG2Loc · 3RScan first 4 scenes, 5 frames | R@25cm,10deg: 0.54 -> 0.61 | Improved |
| add PnP refinement with RoMav2 matching (ablation) | SG2Loc · 3RScan first 4 scenes, 5 frames | R@10cm,5deg: 0.35 -> 0.50 | Improved |
| use proposed particle initialization instead of uniform 3D grid sampling (ablation) | SG2Loc · 3RScan first 4 scenes, 5 frames | R@25cm,10deg: 0.30 -> 0.35 | Improved |
| use adaptive KLD resampling instead of fixed particle count (ablation) | SG2Loc · 3RScan first 4 scenes, 5 frames | R@25cm,10deg: 0.30 -> 0.35 | Improved |
| raycast at max resolution (stride 1) in final pass (ablation) | SG2Loc · 3RScan first 4 scenes, 5 frames | R@25cm,10deg: 0.35 -> 0.33 | No clear effect |
| use longer query sequences for scene retrieval (1 to 50 frames) | SceneGraphLoc extension · 3RScan 50 scenes | Rt@1: 0.69 -> 0.81 | Improved |
| use keyframing with per-keyframe evaluation instead of HLoc | SG2Loc · 3RScan keyframing | median position error (m): 0.63 -> 0.24 | Improved |
| longer sequences (5 to 25 frames) | SG2Loc · 3RScan | joint recall R@0.25m,2deg: 0.34 -> 0.44 | Improved |
| use scene graph map instead of image database (HLoc) | SG2Loc · 3RScan | storage per scene (MB): 57.6 -> 9.8 | Improved |
| use SG2Loc vs HLoc on short sequences | SG2Loc · ScanNet 5 frames | joint recall R@0.25m,2deg: 0.53 -> 0.49 | Hurt |
Closest papers by meaning
- Rooms from Motion: Un-posed Indoor 3D Object Detection as Localization and Mapping (NeurIPS 2025 · similarity 0.89)
- SceneGraphLoc: Cross-Modal Coarse Visual Localization on 3D Scene Graphs (ECCV 2024 · similarity 0.89)
- MeshLoc: Mesh-Based Visual Localization (ECCV 2022 · similarity 0.88)
- EP2P-Loc: End-to-End 3D Point to 2D Pixel Localization for Large-Scale Visual Localization (ICCV 2023 · similarity 0.88)
- Visual Localization Using Imperfect 3D Models From the Internet (CVPR 2023 · similarity 0.88)
- LOCATE 3D: Real-World Object Localization via Self-Supervised Learning in 3D (ICML 2025 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
